Fix normalization handling in MPS/MPO expectations

[gamatos]

This PR fixes normalization issues in tensor network expectation and sweep tracking paths, including avoiding double division for already-normalized MPS inputs and making sweep best-state tracking robust when state is unavailable.

• Fixes expec_mpo so MPS expectation values remain scale-invariant.
• Measures MPS working-data norms via tn_norm while preserving stored exponent state.
• Guards sweep best-state snapshots against missing optimizer state.
• Adds regression coverage for normalized expectations, raw norm measurement, and non-unitary norm-infidelity diagnostics.